Research On Question Answering Technology For Answering History Subject Question

In recent years,artificial intelligence has made a breakthrough in many aspects,so people are paying more and more attention to it.Question answering system(QA)is one of the most important branches of artificial intelligence,and it is also a long-term research target in the field of Natural Language Processing.The existing question answering system can usually be divided into question answering system based on retrieval and question answering system based on knowledge base,both systems need background knowledge before answering questions,but the data in the knowledge base is structured and easy to understand while the other usually contains a large number of Internet text.Several candidate answers will be generated by several related queries,so it needs to calculate the matching degree of each candidate answer to the question which helps to remove the irrelevant candidate answers and obtain the best answer finally.This paper focuses on the related question answering techniques for history subjects,including questions classification,question component extraction,and answer selection.When a question comes,the first step is to make an analysis about the question so as to construct related queries and then some candidate paragraphs will be returned.Finally,calculate the between the sentences in the candidate paragraph and the question to get a short but accurate answer.This paper attempts to apply the method of deep learning to question classification,question component extraction and answer confidence ranking.The content of this paper is as follows:1.We establish data sets of question classification and question component extraction in history subject questions to get the question type and identify the key components of the question.In addition,this paper establishes a data set for answer selection in history subject.2.We construct a question classification model based on deep learning,and use the traditional method SVM for comparison.The experimental results show that deep learning method is superior to the traditional method,and the CNN model achieves the best results with Micro-F1 91.08% and Macro-F1 86.80%.3.We use CRF model and LSTM-CRF model to extract the question component respectively.The experimental results show that the traditional CRF model is better than the deep learning method in the case of small scale corpus with F1 88.51%.4.We construct the framework for answer selection based on deep learning methods,and compare the effect of CNN with LSTM,the answer selection experiments show that the LSTM model outperforms the CNN model.We further test the performace based on different matching degree calculation methods and different loss functions.The experimental results show that using the combination of cosine similarity and Euclidean distance and hinge loss function achieves the best result with MAP 0.4320 and MRR 0.6120.